|
|
|
Data Processing, Algorithm and Modelling
|
Knowledge based object extraction technique
Scenes with Rectangular Shapes in Addition to non Rectangular Shapes
Three different scenes are used in this stage. The first scene contains only one rectangle. The second
one contains two rectangles; the third one contains three rectangles. The edge spread function width
adopted in these scenes is 1.5 pixels.
Figure (3) shows that the entire expected rectangular object was
successfully detected and extracted. One more object (the third object in third experiment) was expected
to be detected; the object fails to satisfy the rules of the rectangular shapes. The reason for that is the
existence of a homogeneous zone near the end of the object, which considers extension for the object.
Figure 3 Results of second type of scenes
Analysis of the results
A twofold criterion is used to analyse the result. It is qualitative in terms of number of extracted
objects compared to the number of actual objects in the scene and the running time and quantitative in
terms of numerical figures, which describe the results. The qualitative result of the experiments shows
that the technique identified all expected rectangles in the scenes. Only one rectangle was expected to
be extracted in the third experiment of the irregular shapes and later this rejection is explained. This
object was not concordant with the rules that define the rectangular shapes. The running time as
component of qualitative assessment points out that the Matlab code is not optimised and it is time
consuming. The quantitative component of the assessment is explored in the following sections.
1. Method of quantitative evaluation
To evaluate quantitative success of the procedure, we have to find truth data to assess the
results. To do that the images that are us ed in the experiments are digitised using on screen digitising.
The digitising is done in Arcview (GIS software). The delineating of boundaries is done in such a way that
it is as close as possible to the reality. It is necessary to say that subjectivity and human delineating
errors are inevitable. The extracted objects are overlaid by the output of on screen digitising. After
overlying the results and the truth data, Arcinfo (GIS software) is used to process the data. Arcinfo is
mainly used to build topology and to calculate the overlapped area. Mainly three quantities are
measured; wrongly predicted parts, parts not predicted, and correctly predicted parts as shown in figure
4. These measurements are done for each object for each test. The chosen quantita tive criterion for
assessment is twofold. The first is related to correctness of extraction and the other relates to the error in
extraction. For the first part the correctness in extraction is considered as the ratio between the correctly
extracted areas to the area of the digitised object. The second part, which is related to the error in
extraction, is divided into two subparts. The first subpart is the ratio of the wrongly predicted area to the.digitised area; the second subpart is the ratio of the not predicted part to the digitised area. It should be
clear that the summation of the three quantities does not produce 100% because these ratios are
normalised by the digitised area. The three quantities: the ratio of correct prediction, the ratio of wrong
prediction and the ratio of the not predicted parts are 97.01,7.56, and 2.99% respectively for the six
experiments.
Figure 4 Assessment procedures
Conclusions and Discussion
This paper has introduced a method for geometrical object parameter estimation of model shapes from
multispectral images. A rectangle model shapes is used. The paper presents a new cost function. The
cost function avoids the problems of using training data in the extraction of agriculture fields. The nature
of the cost function makes the technique independent of assuming any radiometric distribution of the
agriculture field. The output results of several scenes show that the method can detect and extract the
objects with high accuracy. The output of the technique is easy to integrate with GIS. It is produced in two
forms, the first is DXF file and the second is the five parameters of each object.
ACKNOWLEDGEMENTS
This work was done under the Netherlands fellowship program, at ITC, the Netherlands. ITC is deeply
appreciated.
References
- Gile, E., Murray, W., Wright, H. 1997. Practical Optimization, Academic Press, Great Britain
- Van der Heijden, F., 1994. Image Based Measurement Systems: Object Recognition and Parameter
Estimation, Wiley & Sons, Great Britain - ISBN 0-471-95062-9.
- McKeown, j. j., 1990. Introduction to unconstraint optimisation. IOP publishing Ltd., Great Britain.
122p, ISBN 0-7503-0025-6.
- Mulder, N. J. and Fang, L. 1994. Knowledge based image analysis of agricultural fields in remotely
sensed images. Pattern recognition in practice: Vol. IV , pp. 197-211
|
|
|
|
|
|
|
